Principal ML & AI Engineer (Spatial AI & Perception)

• A great work-life balance.
• 30 paid vacation days.
• Hybrid model of work (2 days office and 3 days home office per week).

• On-site Gym and Sauna(For Berlin location), Stay active and unwind with our fully equipped fitness center and relaxing sauna.

• Yoga Room(For Berlin location),  A dedicated space for mindfulness, stretching, and relaxation.

• Work on the development of high-scale services, serving and storing petabytes of data.
• Work with cutting-edge, modern technologies.
• Flexible working hours.
• Competitive salary plus bonus.
• BVG Ticket (For Berlin location).
• German language course (For Germany-based employees).
• Employee wellness programs and life-coaching sessions. 
• A diverse team of fantastic & talented people from 60+ countries worldwide.
• Brown bag talks, team events, BBQ on the rooftop, and more!

Change is HERE. Apply Now!

As part of HERE Technologies employment process, candidates will be required to successfully complete a pre-employment screening process. This offer and any related claims are subject to the successful completion of a pre-employment screening. This will involve employment, education, and criminal verification if applicable.

As ADAS/AD moves towards model-driven intelligence, industry value is extending from map delivery to model training and validation.

It's the growth of HERE's AI-model creation platform that turns maps and drive data into reusable spatial intelligence – powering scalable training, validation, and next generation ADAS/AD performance.

As a Principal Engineer, you will work hands-on with cutting-edge architectures (BEVFormer, BEVFusion, occupancy networks), large-scale distributed training, model quantisation and compression for embedded SOCs, and proprietary datasets at a scale that most organisations simply do not have access to. If you want to push perception AI from research to real-world deployment — this is the role.

• Model Development & Training: Develop and train infrastructure perception models for lane detection, road boundaries, traffic signs/lights, and road surface attributes. Build BEV-based scene representation models (BEVFormer, BEVDet, BEVFusion, or similar), run large-scale distributed training, integrate geospatial priors, and validate synthetic training data for quality and coverage gaps.
• Model Optimisation & Edge Readiness: Own the hands-on model optimisation pipeline for edge deployment, including ONNX export, TensorRT/QNN compilation, operator compatibility checks, graph optimisation, quantisation, compression, and accuracy recovery. Profile and benchmark models on embedded hardware, identifying latency, throughput, memory, and bottleneck issues, and build automated regression pipelines to validate models against release targets.
• Research-to-Pipeline Implementation: Operationalise novel architectures: translate state-of-the-art research (CVPR/NeurIPS papers) into reproducible, scalable training pipelines. Design rigorous experiments, including A/B tests and ablation studies, while evaluating cloud accuracy, deployment readiness, distribution coverage, temporal consistency, spatial coherence, and performance trade-offs across accuracy, latency, memory, and GPU utilisation.
• Collaboration & Quality: Partner with research, simulation, and technical leadership to ensure perception outputs meet downstream training and deployment requirements. Support key architecture decisions around quantisation and precision, design experiments to measure real-world gains from synthetic training, and maintain high standards for code quality, documentation, extensibility, and mentoring the junior engineers as the team grows.

We are looking for a self-directed and pragmatic engineer who can own complex problems end to end, think across the full ML lifecycle from large-scale model training to SoC deployment, and collaborate effectively while maintaining high standards for code quality, documentation, and knowledge sharing.

• 5–7 years in ML/AI with 3+ years in computer vision, perception or deep learning systems
• Proven experience shipping production perception models into real-world systems or large-scale data pipelines
• Hands-on edge deployment experience — optimised and deployed models on embedded hardware or automotive-grade SoCs (not just cloud inference)
• Strong PyTorch expertise, with hands-on experience building custom architectures, loss functions, and training loops, as well as distributed training using DDP/FSDP, mixed precision, and gradient checkpointing.
• Experience with large-scale distributed training, beyond fine-tuning pre-trained models

• Deep Learning & Perception: Experience with BEV or multi-camera transformer architectures such as BEVFormer, BEVDet, BEVFusion, or occupancy networks, along with multi-task learning, scalable training infrastructure, data loading optimisation, augmentation pipelines, and experiment tracking tools such as W&B or MLflow.
• Infrastructure Perception: Experience with structured scene understanding, including road topology, lane geometry, road surface attributes, semantic/instance segmentation of infrastructure elements, and BEV-space, occupancy-grid, or map-aligned representations.
• Model Optimisation & Embedded Deployment (Critical):Hands-on expertise in ONNX and TensorRT/QNN, experience in quantisation workflows such as QAT, PTQ, mixed-precision inference, per-layer sensitivity analysis, and accuracy recovery, along with model compression techniques.Familiarity with automotive SoC platforms such as Qualcomm Snapdragon Ride, NVIDIA Orin, TI TDA4, or comparable embedded accelerators is important, as is experience profiling latency on real hardware, identifying bottlenecks, analysing memory bandwidth, and optimising inference throughput.
• Evaluation & Quality:Ability to evaluate models beyond single metrics, including data coverage, failure modes, and edge cases, while validating deployment readiness through accuracy-latency trade-offs, regression testing across model variants. Ability to reproduce research papers, adapt architectures, and integrate them into scalable, production-ready training loops.